6.2 KiB
6.2 KiB
V1-2.2 owner decisions of 2026-11-01: evidence
Owner brief « V1-2.2 : décisions et poursuite des corrections » on f1bb19b. Owner answers during the
run: threshold 4 CSS px at 320 × 568 (the proposed 6 px refused the validated equine); « the visible
wins » on the editor canvas only. Contract .claude/tasks/contracts/2026-11-01-v1-2-2-decisions.md.
Commits 4f52f5e (lot B), e677150 (R4.1 manifest), 9c606aa (junctions), 96b9978 (visible size,
frame range). Tools .claude/audits/2026-11-01-v1-2-2-decisions-tools/. Details RUST.md §12.12, §12.13.
1. Lot B as approved, isolation re-checked
- Working tree before commit =
.claude/audits/2026-10-30-v1-2-2-lot-b-isolation.patch, identical (index lines aside). - Census
nothing_the_reader_keeps_changes_the_layer_or_the_halo: the reader's 77 allowed attributes (structural ones aside) on the root and a descendant group, 17 descendantstylevalues; layer and halo byte-identical to a plain drawing, root = namespaces + frame. - Browser probe, full-frame opaque magenta rectangle (last drawing child) with 21 candidate styles (15 kept by the reader, 6 removed for a function call), Chromium 156 / Firefox 157 / WebKit 27.2, canvas and PDF: magenta count equal to the unstyled control inside the frame, 13/13 point hits; positive control (rectangle after the layer) 100 % magenta, 0/13 hits.
- Correction to the 2026-10-30 evidence: a descendant
stylemay setpointer-eventsorcursor(the attributes are not kept); it acts on the drawing only, below the layer.
2. References
| Reference | Change |
|---|---|
R3 highlight.json (36), R3 canvas goldens, canvas.*.layers.json |
frozen files unedited; compared under with_isolated_layer / withIsolatedLayer / withDeclaredHitboxes, plus with_canvas_passes / withCanvasPasses for the canvas only (junctions) |
R4.1 pdf/r3/real/manifest.json call 1 svgSha256 |
edited a9a3155c…2c11 → 45fd4aaf…fc7f (e677150); call 2 unchanged; pdf-export.session.test.ts 5/5 |
R4 PDF digests, step6 goldens, N-1 copies, seed review drawings, equine drawing + fingerprint, R3 layer.json, sanitiser.json |
unchanged: git diff --stat f1bb19b..96b9978 over crates/core/tests/{oracles,conformance,fixtures,pdf}, apps/app/tests/fixtures, packages/core-wasm/tests/n-1, supabase/seed lists the manifest's one line only |
3. Junctions, « the visible wins » (canvas only)
- Click spec, 6 projects (Chromium 1280 / 375 touch / 320 touch, Firefox 1280, WebKit 1280 / 375 touch):
99 passed, 0 fixme. On
e677150the junction and disc tests fail in all 6 (12 failures). - Junction test:
c7-t1,l6-s1,t18-l1, fore fetlock, centre + 4 points towards the disc edge, real clicks or taps. Disc test: every twin of main and head, 100 % of samples inside 95 % of the radius record the topmost disc. Zone test: every zone and the trigger, 100 % of samples 2 px inside, away from discs. - WebKit snaps taps and clicks to whole CSS px (measured: tap at 100.1 lands right of an edge at 100.4, click at 100.9 left of it); a real press keeps 1.5 px of margin, below it the browser's hit test answers.
- The five
fixme= one test × five projects; now unmarked in six.
4. Scale bypass and frame range
- Measured editor at 320 × 568 (Chromium): main box 286 × 310, detail 254.9 × 191.1. Equine at 320: junctions 4.97 px, other main points 6.07, head 15.1, smallest zone side 13.1. 6 px refused at 320, 360, 375 px.
sheet_marker_visibility9 tests: equine accepted; horseviewBox× 100 refused with its 19 main markers (marker_too_small); × 1 000 no usable frame; threshold × 1.24 accepted, × 1.25 the 3 junctions, × 1.52 the 13 points; disc vs hitbox; detail box; curves and arcs; every range bound at its limit and past; view_box; drawing transforms never change the measure.sheet_frame13, frame unit test (NaN, ∞).- Rust vs browsers (hitbox layout): equine and thresholds ≤ 8.2e-5 relative, ≤ 0.014 px (Chromium, WebKit); Firefox ≤ 0.4 % (1/60 px app units); far edge 10⁵ / 10⁶ spans drawn raw: 0.14 / 11 px centre error (why the 256 ratio). Painted pixels by span: WebKit nothing from 2⁻⁷, Chromium 0.46 px at 2⁻¹⁴ and nothing at 2⁻²⁰, Firefox exact; floor set at 1 unit (first value 2⁻²⁰ withdrawn on this measure).
5. Gates on the final HEAD 76ead95
- Mutations: on
96b9978,2026-11-01-v1-2-2-mutations-decisions.py29 killed + F5 survived (equivalent: the aspect rule refuses a zero span),2026-10-30-v1-2-2-mutations-three-medium.py26 killed, 6 anchors moved (re-anchored in the new script, killed): 55 killed. N8 (path number bound,76ead95) killed. Total 56 killed, 1 equivalent. pnpm check:allon76ead95passed (01:49 → 01:57): cargo 2 289 passed (worker 49 in Docker), 0 failed, 3 ignored; clippy -D warnings; ts-rs no drift; cargo-deny; RustSec ok; wasm; typecheck; core-wasm 65, supabase 128, app 566; lint; build; e2e:ui 125; app image; worker:test; worker:imagerun --check. Also passed on96b9978.pnpm e2e:shellon the isolated sandbox stack (never DEV): 86 passed on96b9978and on76ead95.- WASM 574 233 B gzip -9 (budget 600 000, margin 25 767 B, 4.3 %); Vite gzip 584.84 kB.
6. Independent reviews (fresh, Opus)
- Verifier on
96b9978: ECARTS(4). D1, D3, D4, D5 met. D2 and D7: §12.12 transform table rows 1 and 3 stale (fixed in76ead95). D6 unverifiable inside its brief (gates above). FLOOR.skip(1)=Iterator::skip(contract clarification added). Former MEDIUMs (1) root property, (2) hitbox stroke, (3) frame and scale bypass, (4) offset slack: closed. MEDIUM path numbers past f32 (fixed76ead95). LOW: lot B commit carried the RUST.md probe and reference passages. - Security on
96b9978: PASS, semgrep 0. MEDIUM path numbers past f32 (fixed76ead95, re-review on the fix: PASS, LOWs informational: rotation reduced in f32, relative commands accumulated in f64, sub-normal arc radius). MEDIUM open, owner decision: zone without interior (zero-winding path, near-collinear polygon) passes the 4 px rule and is hit only on its 2-unit stroke. LOW: twin fully covering an earlier twin (no rule), NaN insideBounds::add(unreachable since the bound), 1e-12 epsilons sound through the span floor, frame walk before the caps (pre-existing).